Apply to ImageNet

[bptripp]

Integrate into a standard network like https://github.com/keras-team/keras/blob/master/keras/applications/resnet50.py

Choose foveation points using a human-foveation prediction network like https://deepgaze.bethgelab.org/

Probably extrapolate images a bit to allow wider range of foveation points without making images too small.